Trajectory
轨迹EssentialA sequence of positions or states and actions over time; also a full recording of one demonstration.
“Trajectory” has two common meanings in embodied AI. In robotics proper, a trajectory is a path plus a timing schedule, specifying which position or joint angles the robot should be at at each moment; the textbook Modern Robotics defines it as “a path plus time scaling,” and a controller's job is to track this trajectory. In reinforcement learning and in datasets, a trajectory refers to the sequence of states and actions produced by an agent interacting with the environment, τ = (s0, a0, s1, a1, …); one recorded teleoperation demonstration is one trajectory, and datasets such as Open X-Embodiment measure their scale in number of trajectories. A short segment of future actions a model outputs at once, an action chunk, is also often loosely called a short trajectory.
ExampleThe 150 end-effector poses and gripper-opening values recorded at 50 Hz over the 3 seconds a robot arm takes to pick up a cup from a table make up one trajectory.
- Related
- Episode · Trajectory Planning · Trajectory Tracking · Demonstration Data · Action Chunking · Path Planning
- Sources
- OpenAI Spinning Up: Key Concepts in RL(Trajectories)
Lynch & Park, Modern Robotics, Ch.9 Trajectory Generation(预印本 PDF) (Chinese)